The Conditional CAPM and the Cross Section of Expected returns

Introduction
This lecture note describes the progression of the empirical asset pricing literature begun by Fama and French in their 1992 paper. The main papers discussed
are
1. Fama French 1992 The

Cross{Section of Expected Stock Returns

Size and Book to Market capture the cross{sectional variation in expected stock returns associated with beta. The relationship between
expected return and the component of beta unrelated to size is at,
even when beta is the only explanatory variable in the regression.

2. Fama French 1993

Bonds

Common Risk Factors in the Returns on Stocks and

Denes the basic FF 3{factor model.

3. Fama French 1996 Multifactor

Explanations of Asset Pricing Anomalies

Explores whether anomalies can be explained by the FF 3{factor

model.

Along the way we will relate these papers to, among others
1. Kothari Shanken and Sloan (1995)
2. Daniel and Titman (1997)

FF (1992)

Note Banz (1981) size eect (low is good)

Note Book/Market eect (high is good)

leverage apart from beta (taxes?)

Basu (1983) E/P ratio. But Ball argues that this a proxy for risk.
1

BJS and FM nd pre 1969 positive relationship between leverage returns

and 

Results

Relationship disappears in 1963{1990 period.

Also weak in 1941{1990 period.

But size, Book/market, leverage, E/P all load.

Most basic prediction of CAPM not found in data.

Data

NYSE NASDAQ AMEX.

Compustat needed for book values

Match accounting data for all scal year ends in calendar year t

1 with

returns for July of year t to June of t +1. The 6 month gap is conservative.

Use market equity in December of t 1 to compute book/market, leverage,

E/P rations for t

1 and use

Market equity in June of year t to measure size

Estimating betas

FM approach

Form size portfolios

Then subdivide based on pre ranking betas for individual stocks

this yields 100 portfolios

Table 1 provides summary statistics.

{

note that beta decreases with size

note that even after accounting for size, there is substantial variation
in beta within size deciles

note that the aforementioned variation in beta occurs even though

size is essentially constant across the beta-sorted deciles (within the
size deciles).

FM Regressions: Regress sorted portfolios on beta, size, book/market, and other

stu. Results are in table 3.

Note rst that when only beta is included in the regression, the market
risk premium is only .15 percent per month(incredibly low, even with
a potential errors in variables problem | the betas are measured with
error), which is statistically insignicantly dierent from zero.

Size and book to market load in the appropriate ways. Size is negatively
associated with expected return, and book/market is positively related.

One explanation for the book/market premium the FF have is nancial

distress risk premium that is not captured by beta.

Note that (p. 440) in the 1941{1965 period, which is most of the BJS and
FM sample, the simple relationship between beta and average return is
present, although it disappears when controlling for size.

Leverage. Use 2 variables, book assets / market equity and book assets /
book equity.
{

opposite slopes but close in absolute value at about .5

explanation: dierence between market and book leverage helps explain average returns

E/P ratio: Ball (1978) say this should be related to residual risk if asset
pricing model is incorrectly specied.
{

O.k. for positive earnings, what about negative earnings?

negative earnings not a proxy for forecasts

so have a dummy for negative earnings rms

result is that when earnings are positive, expected returns on average

are 4.72 percent per month (huge).

But adding size and book/market kills eect.

3

FF 1993

Denes the basic FF 3{factor model. Extends asset pricing tests of FF 1992 in
three ways:
1. tries to explain both stock and bond returns
2. add term structure variables to help explain bond returns (in addition to
SML and B/M
3. uses a slightly dierent approach to test - runs time series regressions on
the factors where the slopes are the traditional factor loadings (ask why
this is so in class)1 Then they do a test ala GRS (1989). This is a big
improvement.

What's nice about the approach is that from the time series regression, the R2 has meaning (how much cross-sectional variation explained).

Main Results

For stocks, portfolios constructed to mimic size and B/M capture strong
common variation in returns, no matter what else is in the time series
regressions.

These factor mimicking portfolios proxy for some unknown factors.

The intercepts are close to zero (that doesn't mean that some anomalies
don't remain anomalies vs. the FF 3 factor model - that sort of alternative
is not explored here.

Although the size and B/M portfolios explain cross sectional variation in
stock returns, they don't explain the equity premium. This is left to the
market factor. Essentially, all of the stock portfolios produce market betas
close to one.

So, the equity premium is left unexplained. The FF model just says that
the dierence between the returns on individual stocks and bills can be

1 The

answer is that size and book to market don't mean anything for bonds.

explained by the market risk premium. Where that premium comes from
is left unexplored.

For bonds, the mimicking portfolios for the 2 term structure factors - a
term premium and a default premium (as in CRR 1985) capture variations
in the returns on bonds.

The term structure factors also \explain" the average returns on bonds,
but the average premia for the term structure factors are about zero, so
that all bond portfolios statistically speaking, have the same returns (this
is bad for the model).

The stock returns are explained by 3 factors, the bonds by 2, but if you
only include the terms structure factors and try to explain stock returns,
you're not unsuccessful.

The bond returns can be explained by the 3 stock factors, but these are
driven out when the term structure factors are put back, the stock factors
don't load except for with risky bonds.

2.1

Factors

Bond factors
1. Term is the dierence between the monthly long term government return
and the one month t{bill rate
2. DEF is the dierence between a portfolio of long{term corporates and the
long{term government return
Stock Factors:
Size and book/market factors. This is how it works.

In June of each year, all NYSE stocks on CRSP are ranked on size.

The median NYSE size is used to split up NYSE, NASDAQ, and AMEX
stocks into S and L categories.

Also, NYSE, Amex, and NASDAQ stocks are broken into 3 B/M groups
5

Low 30%

Medium 40%

High 30%

Details are on PP 8{9 of paper

The splits are arbitrary

Then, construct 6 portfolios: (S/L, S/M, S/H, B/L, B/M, B/H) from the
intersections of the above portfolios.

Monthly value weighted returns are calculated from July of that year to
June of the next and then the portfolios are reformed next June.

Now, the portfolio SMB is simply the dierence between the simple average of the returns on the 3 S portfolios and the 3 B portfolios.

HML is dened similarly, but the medium book/market stocks are not
included in constructing this factor.

The correlation between these two factors in the 1963{1991 period is -0.08!

Finally, the market factor is simply the excess return on the market over
the t{bill rate.

The returns to be explained. Bonds.

Excess returns on 2 government and 5 corporate bond portfolios.

Governments: 1{5 years; 6{10 years

Corporate: Aaa, Aa, A, Baa, LG

The returns to be explained. Stocks.

Excess returns on 25 portfolios formed on size and book/market equity.

Later on, use portfolios formed on E/P and Div/P to check robustness

Sort by size and (independently) B/M

The sorts use NYSE breakpoints.

6

Then, combine into the 25 portfolios.

this is done in June of each year and as above, the portfolios live for a
year before the sort is done again.

Table 1 (p. 11) gives descriptive statistics.

Return Characteristics: Dependent Returns

Wide range of excess returns, from 0.32% to 1.05% per month.

Negative relation between size and average return

a stronger positive relation between B/M and return

Most of the 10 portfolios in the bottom two B/M quintiles are not statistically signicantly dierent from zero.

For bonds, average excess returns are puny, less than 0.15% per month.

Government returns don't increase with maturity

Corporates have higher returns than governments

Low grade bonds have higher returns than high grade

Return Characteristics: Factors

Market risk premium in sample averages 0.43% per month

SMB: 0.27% per month (slopes will be high though for our portfolios)

HML: 0.40% per month

TERM: 0.06% per month (statistically insignicant)

DEF 0.02% per month (statistically insignicant)

2.2

Results

Table 3: Regressors are bond factors only.

Bond factors explain stock returns when stock factors are not included,
but regression R2 are low.
7

Bond market factors explain bond returns very well.

TERM slopes are higher for long{term securities | one would expect log
term securities to be more sensitive than short{term securities to shifts in
interest rates.

DEF seems to capture a common default risk in returns (bigger for stocks
than risky bonds than treasuries)

one problem is that the risk premia on the bond factors is tiny, and is not
likely to be able to explain the size eect for stocks

When stock factors are added to the regressions the loadings on the bond
factors (for stocks) diminish or vanish.

Table 4: CAPM

Regressions t reasonably well for stocks, but not for bonds.

bond regressions at least make sense (of course, the stock regressions do
too) - corporates are more risky than governments.

Table 5: Regressors are stock factors but not market factor.

bond R2 are zero

Stock R2 are lower than in table 4!

Table 6: 3 Factor FF Model

good t! Note that the R2 here are not the .5 R2 that JW are talking
about in the cross{sectional regressions like FM (1973). The R2 reported
here are for time series regressions.

The factors aren't independent. The correlation between the market risk
premium and SML and HML are .32 and -.38. That's not that high.

Note pattern in the betas. There's not much cross{sectional variation in

beta any more. Beta behaves a lot like a (changing) intercept term.

That is, it looks not like the excess return on a portfolio in a particular
period equals the excess return on the market in that period plus the size
and book/market loadings (where there is cross{sectional variation. That
is saying that stocks move with the market really well, but the extent to
which there is cross{sectional variation in stock returns, it is driven by
cross{sectional variation in size and book/market factors.

In my mind, this is the most important result in the paper.

This

is the only way to get FF 1993 and FF 1992 to tell a consistent and
powerful story. I think that the paper essentially makes this point in the
introduction, but I would have liked to see it made here again.

The bond t is much less good.

Table 7a: 5 Factor FF Model (Kitchen sink) dependent variable is stocks

Fit is still good | bond factors don't matter much

Table 7b: 5 Factor FF Model (Kitchen sink) dependent variable is bonds

Fit is good | stock factors matter (but not much)

Table 8: RMO stu. RMO is an orthogonalized market factor | regress RM

Rf on other factors.

Table 8a Now term structure variables load for stocks

Table 8b and beta loads for bonds

Table 9: Intercepts

9a. 2 factor bond model produces large positive intercepts

9a. CAPM produces a lot of cross{sectional variation in intercept

9a. 2 factor stock model produces large positive intercepts

9a. 3 and 5 factor models do much better but 5 factor intercepts are
almost same as 3 factor.

GRS Test (the formal test on intercepts)

reject at the .95 level for the 3{factor model

9

2.3

Diagnostics

Regression residuals appear to be unpredictable.

January eect largely explained. Essentially, there's a seasonality in the

size risk factor. This is contrary to results in CK (1986, 1988 1993) where
curve tting for factors (which allows time variation in factor premia as
well) cannot explain the January eect.

Portfolios formed on E/P

{

conrm Basu that E/P not explained by the market model

FF 3{factor model looks good, but where's the GRS test?

Portfolios formed on D/P

{

see table 11 - a lot of cross{sectional variation explained, but where's

the F-test?

FF 1996

The basic model:

E [Ri ]

Rf = bi (E [RM ]

Rf ) + si E [SMB ] + hi E [HML]

The basic time{series regression:

Rit

Rf = ai + bi (RMt

Rf ) + si SMBt + hi HMLt

with zero intercept.

FF 1995: Slopes on HML and Book/Market proxy for relative distress

(also in 1992 paper)

FF 1994: Use basic model to explain industry returns.

This paper:

Shows that the 3{factor model captures returns to portfolios formed on

1. E/P
10

2. C/P
3. Sales growth

Low E/P. C/P, and high sales growth are typical of strong rms (which
have a negative loading on HML | the average HML return in strongly
positive | 6% premium.

Also, the model explains long{term reversals (see e.g. DeBondt and Thaler
1985, FF 1988) but not 3{12 month momentum (JT 1993)

Given results the authors claim that the 3{factor model is an equilibrium asset
pricing model but that

Their \factors" are actually factor{mimicking portfolios.

There has been some skepticism

1. Kothari Shanken Sloan (1995) say that the HML factor is overstated due
to survivorship bias on compustat. A disproportionate number of high
book/market rms actually survive distress.
2. There may also be a problem with data snooping Lo MacKinlay (1989)
where the results only hold in sample.
3. There may also be some irrationality at work (Lakonishok, Schleifer, Vishny
(1994)) where overreaction leads to underpricing of distressed stocks.
The paper discusses these competing stories in section VI.
3.1

Tests on the 25 Size-BE/ME Portfolios

Table 1

R2 high

What about cross{sectional variation in beta?

Test interpretation? (This is given in my notes on FF 1993, but I want to

ask the class about it. It is a very important point.

GRS test: reject. FF spin: intercepts are measured very accurately because model explains so much cross{sectional variation.
11

3.2

LSV Deciles

Sort on
1. B/M
2. E/P
3. Cash/P
4. 5 year sales rank
Table 2, 3, 4, 5 shows that the 3 factor model explains this cross{section of
average returns.

for example: C/P deciles. High C/P are high SMB and high HML. High
C/P stocks are relatively distressed (interpretation).

sales rank explained less well (intercepts have a pattern)

GRS: don't reject!

tables 4, 5: double{sort portfolios, similar results

3.3

Long{term winners become losers

DeBondt and Thaler (1985). Portfolios are formed based on long{term (3 to 5

year) past returns). This is explained.

past losers load on SMB and HML. They behave like small distressed
stocks, which have high future returns.

Table 7: can't explain momentum.

3.4

Exploring 3{factor models

Tests: essectially, of M, S, H, L, can any three of these explain the fourth? Put
another way, loosely speaking, are thee factors enough? The answer appears to
be yes.

Note that LSV proxies cannot replace L or H in the three factor model
12

3.5

Can Loadings on 3 factors explain the cross{section

of average returns?

3 Stories
1. Asset pricing is rational, and that the equilibrium model is a 3 factor
ICAPM or APT
2. Irrationality (overreaction) causes a high premium for relative distress.
(but table 9 shows that relative distress looks like a factor, and has a big
enough standard deviation that there's no arbitrage opportunity.)
3. CAPM holds but is spurriously rejected either because
(a) Survivor bias in returns used to test the model KSS (1995)
(b) Data snooping (counterarguments: holds in 1941-1962 and internationally)
(c) Bad proxies (JW)

Daniel and Titman 1997

4.1

Introduction

Disagreement about the high discount rates for rms with

1. small size
2. high book/market

FF have made the case that it's systematic risk | that B/M proxies for
a distress factor. For example, distressed rms are sensitive to uncertain
market conditions, credit, to name one.

LSV claim that high B/M is due to market overreaciton | bad past performance leads these rms to be irrationally "undervalued." In addition,
good rms may ve overvalued (also due to investor naivitee).

FF 1993 tests:
13

prices of high B/M rms move together

B/M factor explains cross section of expected returns

LSV then say that the factor risk premium on the B/M factor is too big,
and it's not correlated with macro factors anyway.

DT say that LSV results are not inconsistent with a Merton{type factor
model where there is a priced factor (or os) associated with changes in
the investment opportunity set that is orthogonal to the market return
(continuous time conditional CAPM). However, this is a dicult thing to
nd.

So instead, what DT do is look

1. are there pervasive factors associated with size and B/M
2. Are there risk premia associated with the factors?

The results are

1. No discernable risk factor associated with high B/M
2. No return premium associated with any of the 3 factors. (strange)
3. Although high B/M stocks covary, they covaried just as much before
they became distressed | the covariance is due to something else,
like industry, or some other rm characteristics.
4. Perhaps the main result is that portfolios with similar characteristics,
but dierent FF factor loadings, have dierent returns, and they
don't.

4.2

Summary of return patterns of size and B/M portfolios

Take the 25 FF (1993) portfolios over 7/63{12/93.

Panel A: high B/M quintiles outperform low B/M quintiles after controlling for size. After controlling for B/M, size is a large rm anomaly |
large rms have signicantly lower returns than the rest.

Panels B and C: Size and January are the same eect.

14

4.3

A model of the return generating process

3 models
1. Null: FF 3 factor model
2. Alternative with stable factor structure, but expected returns are determined by rm's loading on factors with time{varying return premia.
3. Firm characteristics, not factor loadings, determine expected returns.
Model 1: standard
Model 2: Time{varying factor risk premia.

factors don't change as rms become distressed

However, distressed rms will have loadings on factors with bad outcomes
in the past.

For example, if oil realizations bad, high B/M (and small size) rms will
be more weighted to oil stocks.

Also, changes in the time{varying. risk premia are negatively correlated

with past performance of rms loading on that factor.

Finally, there exists an observable variable it that is mean reverting and
innovations in  are negatively correlated with past returns. So across
rms,  should be correlated with the rms loading on the distressed
factor.

That is, high  stocks have factor loadings on factors with high s so that
they get high expected returns.

Model 3: Characteristics{based model

Time{invariant J factor structure

rit = Erit +

with fjt  N (0; 1).

15

X
J

ij fjt + it

But factor loadings do not describe expected returns

Erit = a + b1 it

where  is some slowly{decaying rm attribute. Innovations in  are negatively correlated with returns on the stock, but  is not directly related
to loadings on the distress factors.

As in model 2, innovations in  are negatively correlated with returns

on the stock, but  is not directly related to the loadings on the distress
factors.

What is unique about model 3: rms exist that load on the distress factors but are not themselves distressed, and therefore have a low  and
commensurately low return.

If model 3 is true, there may be some stocks which despite high loadings
on an oil factor that has had a series of bad realizations, are still not
distressed.

Model 2 says that they should still earn the distress premium.

Model 3 says that they shouldn't | it's all about 

Empirical implications of the models

FF: (1) high B/M stocks covary and (2) they have high returns.

The conclusion from (1) and (2) is that there is a B/M factor that has a
high premium.

Models 2 and 3: this conclusion need not follow from the evidence.

The reason is that a string of negative factor realizations will cause rms
that load on that factor to become distressed. Distressed rms covary, but
not because of a distress factor.

To discriminate between model 1 and models 2 and 3, see if the return

s.d. of a portfolio of rms increases if they all simultaneously become
distressed.

16

If the factor structure is stable and there is no separate distress factor,

the return s.d. should be constant.

What would time{varying betas on the distress factor do to this test?

Finally, model 3 also indicates that these observations don't necessarily

mean that returns determined by factor loadings.

In the characteristics based model, high returns are earned by all distressed
rms, whether or not they load on the distress factor.

Note that the stability of the covariance matrix is important for testing
pricing aspects of characteristics based model.

4.4

Covariation of stocks wih similar characteristics

Characteristics model diers from model 1 in 2 ways

1. Characteristics model has no distress factor
2. Average returns determined by characteristics, not factor loadings

(1) is important because common variation among value and growth stocks
has been interpreted as evidence of a distress factor.

The reasoning is: if you select 1000 stocks and go long, and 1000 stocks
and go short, the resulting portfolio has low return variance.

But HML portfolio has a much higher standard deviation than that.

In contrast, the characteristics model assumes that this common variation

arises because the HML portfolio changes, and always has stocks that
move together in it. They are just dierent stocks at dierent points in
time.

The Portfolio returns

If model 1 characterizes data, stocks should have higher covariances when

they are in the high B/M portfolio than when they are not.

If 2 or 3 true, covariances constant over time.

17

So DT form the 6 portfolios based on HML and BS interaction and then

form SMB and HML.

Then calculated pre formation and post formation return standard deviations in 10 years centering around portfolio formation date. Keep portfolio
weights constant.

Before looking at these, look at average returns of stocks in the HML

portfolios prior to formation date.

Pre formation returns negative (no surprise)

s.d.'s should go up, but it doesn't. Of course, this is because all these rms
experience negative returns at the time leading up to portfolio formation.

But this is also true as the stocks go forward!

Cross sectional test of factor model

If factor model correct, high B/M factor loading = high E[return]

need portfolios with low correlation between factor loadings and characteristics (high B/M ratios but low loadings on HML factor).

So, rst rank rms on B/M into High, medium, and low and on size.

Then, place these 9 portfolios into smaller ones, based on HML factor
loadings. That is, stocks have same size, and B/M but dierent factor
loadings.

Do they have the same returns (as characteristics based model predicts)
or dierent ones (like a factor model predicts)?

Table 3: mean excess returns for the 45 portfolios. Weak relation between
returns and factor loadings for larger stocks only.

Maybe that is true because in sorting on HML factor loading, we are

picking up some variation in book/market. This is borne out in table 4.

Table 5 shows that there is considerable dispersion in post{formation factor loadings but when you regress post formation excess returns for each
of the 45 portfolios in the factors you get
18

HML coecients are dierent for dierent B/M groups

But WITHIN a B/M size group, the sort of pre formation HML factor loadings produces a monotonic ordering in post formation factor
loadings.

Models 1 and 2 say intercepts should be zero in factor regressions.

Model 3 says alphas of low factor loading portfolios should be positive
and high loadings should have negative intercepts. This is true (table
5 column 5). There is a formal test in table 6.

Table 7,8 tests to see if SMB and MKT factors are priced after controlling
for characteristics.

Size is marginally signicant.

Mkt: reproduce FF 92 - after controlling for size, relationship for beta is

slightly negative.

19